Silicon Wafer Defect Detection Algorithm Based On Deep Learning

Silicon wafers are the main raw material for transistors,solar cells and integrated circuits.It is important to detect the defects of silicon wafer quickly and accurately.The detection method based on machine vision has higher precision and lower cost than traditional methods such as artificial detection,so it is gradually widely used.In particular,the development of deep learning has fur-ther improved the accuracy and speed of target(defect)detection algorithm.However,in the process of defect detection of silicon wafers,there are some problems,such as small defect sample data set,unbalanced defect sample data,and easy to miss the detection of micro-defects such as holes.In this paper,the problems mentioned above are studied in depth.The main research contents include:1.The transfer pre-training method based on the generative adversarial networks is proposed to solve the problem that the detection model is easy to overfitting when the defect sample set is small.Firstly,the generative adversarial networks is used to learn the data set of silicon wafer,and a pre-training data set is made by using the generating model.In the pre-training data set,the defect detection network is pre-trained for migration,which accelerates the convergence of the network and improves the detection effect of the model.Secondly,the defect data set is expanded by data augmentation method,and the number of different samples is balanced,which makes the detection network have better detection effect for various defects and improves the generalization ability of the network.Experiments are carried out to analyze the effect of generative countermeasure network and data augmentation method on improving the accuracy of defect detection algorithm.The experimental results show that the mAP of the detection network obtained by this method is 50%(from 0.44 to 0.66)higher than that in the original case,especially for the defect categories with fewer samples,the detection accuracy is improved significantly(the hole AP is increased from 0.21 to 0.56).2.A small target detection algorithm based on improved YOLO network is constructed.Firstly,the improved network uses the method of constructing feature pyramid model to predict multi-scale,and detects defective targets from three different scale feature maps,so as to improve the detection accuracy of small targets.Secondly,residual network is used to extract image features,and SENet module is added to calibrate feature channels.By adaptively calibrating the correlation between different channels of convolution network,the detection effect of the model is improved and the convergence of the network is accelerated.The experimental results show that the improved detection algorithm achieves 0.94 mAP for defect detection and 98.67%accuracy on test set,while the AP for small defects such as holes achieves 0.91,which is 62%higher than the original network(from 0.56 to 0.91).